Follow the hash, not the hype.
NVIDIA's FY2025 Q2 report, filed for the period ending July 28, 2024, shows revenue at $96.2 billion, a 106% year-on-year increase. The market reads this as a clear signal of AI dominance. I read it as a stress test on a supply chain that is operating at 100% capacity. The headline numbers are real. The underlying fragility is the story that matters.
This is not a critique of the P&L. It is an audit of the physical constraints that sit behind every line item. When a company reports a 74.5% gross margin and $213.4 billion in free cash flow, the question is not whether the demand is real. The question is whether the supply chain can sustain the trajectory without breaking.
The answer is more nuanced than the stock price suggests.
Context: The Fabless Monolith
NVIDIA operates as a fabless designer with a system-level solution stack. It does not own fabs. It owns the architecture, the interconnect, and the software ecosystem. This is a deliberate design choice. It keeps depreciation low and return on invested capital high. The current product line, Hopper (H100/H200), is built on TSMC's 4N process. The next generation, Blackwell (B100/B200), moves to TSMC's 4NP node with CoWoS-L packaging. The roadmap is aggressive: Blackwell Ultra in 2025, Vera Rubin in 2026 on TSMC's N3 process.
The competitive position appears unassailable. NVIDIA controls over 90% of the AI training GPU market. The CUDA software stack is a moat that competitors cannot easily cross. AMD's MI300 series is close on hardware specs but far behind on ecosystem. Intel is effectively a non-factor. The product iteration cycle has been shortened from two years to one, a move that signals deep technical reserves and an intent to keep the pressure on.
But the numbers reveal a dependency that is not priced into the narrative. The hyperscaler segment accounts for roughly 54% of revenue. Microsoft alone represents 15-20%. This concentration is not a flaw in execution. It is a structural reality of the AI build-out. The demand is real. The concentration risk is also real.
Core: The Ledger Behind the Headlines
Let me walk through the technical details that the earnings call glossed over.
The CoWoS Bottleneck
TSMC's CoWoS capacity is the single largest constraint on NVIDIA's ability to ship. Utilization is at approximately 100%. This is not an estimate. It is a physical reality. NVIDIA is the largest consumer of CoWoS, taking up over 60% of TSMC's output. Any delay in TSMC's expansion plan directly caps NVIDIA's revenue growth.
The Q3 guidance shows a gross margin of 73.5% to 74.5%, slightly below Q2's actual 75%. This is a tell. It implies that Blackwell's initial yield ramp is costing more than expected, or that CoWoS capacity is being rationed at a premium. Either way, the margin compression is a direct result of the supply chain's inability to scale at the same pace as demand.
The HBM Dependency
High Bandwidth Memory is the other critical input. SK Hynix, Samsung, and Micron supply HBM3E, and the market is undersupplied. NVIDIA has prepaid to lock in capacity, but this is not a guarantee. It is a hedge. The prepayments also explain the gap between net income and free cash flow. The cash is being spent before the product ships. This is a smart move, but it is also a signal of how tight the market is.
The Hidden CapEx
NVIDIA's own capital expenditure is modest, around 5-8% of revenue. But the effective capital intensity is much higher when you include the prepayments to TSMC and memory suppliers. This is an off-balance-sheet commitment that will show up as inventory and receivables in future quarters. Investors who focus only on the reported CapEx are missing the real cash flow picture.
The Yield Question
Blackwell's initial yields are estimated at 60-70%. This is below the healthy 80%+ level that TSMC achieves on mature nodes. The 4NP process and CoWoS-L packaging are new, and the learning curve is steep. The company expects yields to improve by mid-2025. That is an assumption, not a certainty. If yields do not improve as expected, the margin pressure will persist longer than the market anticipates.
The Geopolitical Overlay
Export controls have reduced China's contribution from 20% of revenue to roughly 10%. This is a manageable hit, but it is not zero. The company is exploring sovereign AI deals in the Middle East and Southeast Asia to fill the gap. These deals are slower to close and come with their own regulatory risks. The US government has already signaled that it may tighten restrictions on exports to the Middle East. This is a tail risk that is not fully priced in.
Based on my audit experience, the pattern is clear. The company is executing well within its control. The risks are external. The supply chain is the binding constraint. The geopolitical environment is the wildcard.
Contrarian: What the Bulls Get Right
The bulls are not wrong about the demand. AI compute is a structural shift, not a cyclical boom. The capex commitments from Microsoft, Google, Amazon, and Meta exceed $200 billion in 2024. This is real money being spent on real infrastructure. The inference workload is just beginning to scale, and it is likely to exceed training demand by 2025. This is a multi-year growth story.
The bulls are also right about the ecosystem. CUDA is a lock-in that cannot be replicated quickly. Even if cloud providers like Amazon and Google develop their own ASICs for inference, the training market will remain NVIDIA's for the next 3-5 years. The software moat is stronger than the hardware advantage.
The valuation, at roughly 60x trailing earnings, is high but not unreasonable if the growth continues at a 50%+ CAGR. The PEG ratio of 1.5x is within the range of historical high-growth tech stocks. The market is paying for quality, and NVIDIA has delivered on its promises.
What the bulls miss is the timing risk. The supply chain constraints are not permanent, but they are real for the next two to three quarters. If CoWoS expansion slips or HBM supply tightens further, the revenue growth will decelerate faster than the guidance suggests. The margin guidance already hints at this. The market is not listening.
Takeaway: The Physical Reality Check
The numbers are impressive. The technology is best-in-class. The moat is real. But the ledger does not lie. NVIDIA's growth is constrained by the physical capacity of its suppliers. The company has done everything right to secure that capacity, but it cannot control the execution of TSMC or the memory makers.
On-chain evidence never sleeps. The same principle applies to physical supply chains. The data is available. The question is whether the market is reading it correctly.
Check the multisig. Always. The next earnings report will show whether the supply chain held up. If the margin compression deepens, the stock will face a correction. If the yield improves and the capacity expands, the growth story continues. The data will tell the story. The hype will not.
The real test is not the next quarter. It is the next two years. The physical constraints will be the deciding factor. The market is pricing in perfection. Perfection is rare in manufacturing.
